Abstract
BACKGROUND/OBJECTIVE: We combined cranial accelerometry, a device-based approach to large vessel occlusion (LVO) prediction, with neurological examination findings to determine if this improves diagnostic accuracy compared to either alone. METHODS: Cranial accelerometry recordings and NIHSS scores were obtained during stroke codes and thrombectomy transfers at an academic medical center using convenience sampling. The reference standard was discharge diagnosis of LVO stroke. We compared accuracy statistics between machine learning models trained using cranial accelerometry alone, with asymmetric arm weakness added, with NIHSS scores added, and retrospective examination only LVO prediction scales. An exploratory analysis required asymmetric arm weakness prior to model training or scale testing. RESULTS: Of 68 patients, there were 23 LVO strokes. Cranial accelerometry was 65% sensitive (95% CI 43-84%) and 87% specific (95% CI 73-95%). Adding asymmetric arm weakness increased specificity to 91% (95% CI 79-98%). Adding asymmetric arm weakness and the NIHSS increased sensitivity to 74% (95% CI 52-90%) and decreased specificity to 89% (95% CI 76-96%). LVO prediction scales had wide sensitivity and specificity ranges. The exploratory analysis improved sensitivity to 91% (95% CI 72-99%) and specificity to 93% (95% CI 92-99%) with only three false positives and two false negatives. CONCLUSIONS: Cranial accelerometry models are improved by various additions of asymmetric arm weakness and the NIHSS. An exploratory analysis requiring asymmetric arm weakness prior to cranial accelerometry model training minimized false positives and negatives.